Pith. sign in

REVIEW 5 major objections 4 minor 21 references

Hybrid physics-LSTM model tops baselines for lane-change intent

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 13:26 UTC pith:FKKOHUPA

load-bearing objection The hybrid idea is plausible but the paper's reported numbers are inconsistent across three places and the history-window selection leaks test information, so the central empirical claim is not supportable as written. the 5 major comments →

arxiv 2512.24075 v5 pith:FKKOHUPA submitted 2025-12-30 cs.LG

Evolutionary Physics-Informed Temporal Fusion for Lane-Change Intention Prediction

Jiazhao Shi , Qiyang Xie , Ziyu Wang , Dongxu Zhang , Yichen Lin , Di Zhu , Chen Xie , Ziwei Wang
show 3 more authors
Haoyun Zhang Enliang Li Zetong Guan
This is my paper
classification cs.LG
keywords lane-change intention predictionphysics-informed AIBi-LSTMLightGBMtemporal fusionclass imbalancehighDexiD
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that combining learned temporal representations from a bidirectional LSTM with handcrafted physics-inspired interaction features in a gradient-boosted tree classifier yields consistently higher Macro F1 for three-class lane-change intention prediction (left, right, none) than either component alone, on two drone-based highway datasets at 1, 2, and 3 second horizons. The advantage is largest in ramp-rich merging scenarios, where interactions are dense and longer history windows help. If true, it suggests that physics-informed feature engineering and deep sequence learning are complementary rather than competing approaches for early maneuver prediction.

Core claim

On its own terms, the paper establishes that the hybrid TPI-AI framework—concatenating a pooled two-layer Bi-LSTM embedding of raw trajectory history with kinematics, safety (DHW, THW, TTC), and interaction features, then classifying with LightGBM under imbalance-aware resampling/weighting and threshold calibration—outperforms standalone LightGBM and Bi-LSTM baselines in Macro F1 across all tested horizons and both datasets. Concretely, it reports Macro F1 improvements over the best baseline on highD from 0.9359 to 0.9562 at 1s, 0.8916 to 0.9124 at 2s, and 0.8115 to 0.8345 at 3s; on exiD from 0.9004 to 0.9247 at 1s, 0.8084 to 0.8197 at 2s, and 0.7127 to 0.7605 at 3s. The paper also shows tha

What carries the argument

The central mechanism is feature fusion: a two-layer bidirectional LSTM encoder with 256 hidden units per direction pools its final hidden states into a fixed embedding, which is concatenated with a large set of physics-guided handcrafted features (243-264 variables), and this fused vector is fed to a LightGBM gradient-boosted tree classifier. The Bi-LSTM captures preparatory temporal dynamics; the handcrafted features encode compact risk and feasibility surrogates (TTC, headway, safe-gap indicators, lane-advantage indices); LightGBM provides robust tabular decision boundaries and handles class imbalance via histogram-based splits, leaf-wise growth, GOSS and EFB. The imbalance pipeline (SMOT

Load-bearing premise

The history window length W was selected using the held-out test locations, so the reported Macro F1 scores for the chosen windows likely overstate generalization; if W were fixed a priori or chosen by nested cross-validation, the hybrid's advantage over baselines could shrink.

What would settle it

A replication on the same datasets with nested cross-validation for window selection (or a fixed window per horizon) that yields Macro F1 gaps smaller than the reported ~0.02 points on highD or fails to beat the LightGBM baseline on exiD at 2-3s would refute the central claim of consistent hybrid superiority.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If correct, hybrid physics-informed plus learned temporal fusion is a robust recipe for early intention prediction in heterogeneous highway traffic, not just lane changes specifically.
  • The pronounced exiD gains suggest ramp/merge scenarios especially benefit from temporal physical descriptors, pointing to environment-specific model configuration.
  • The systematic degradation with horizon confirms an anticipation-reliability trade-off; deployments must set operational horizons accordingly.
  • The large macro-F1 gains from imbalance handling (24 points on average on highD) imply that class-imbalance mitigation is a primary lever for minority maneuver detection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's abstract and full text differ in feature nomenclature ('evolutionary temporal descriptors' vs. 'physics-guided handcrafted features'); if the temporal descriptors are the actual contribution, the body under-specifies them, and a replication would need to clarify which feature set drives the gains.
  • A clean ablation—removing physics features, removing Bi-LSTM embeddings, and removing the imbalance pipeline—would test whether the fusion is truly additive or whether the imbalance pipeline alone explains most of the gain.
  • The location-based split prevents trajectory overlap but does not by itself guarantee that selected history windows are unbiased; windows were chosen on held-out test locations, so the reported scores are optimistic. Nested cross-validation or a three-way split would test robustness.
  • The approach transfers naturally to other maneuver-intention tasks (turn, cut-in, U-turn) and to multimodal inputs (camera, radar, V2X), which the paper lists as future work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes TPI-AI, a hybrid framework that concatenates Bi-LSTM temporal embeddings of raw trajectory histories with handcrafted physics-informed features and feeds the fused representation to a LightGBM classifier for three-class lane-change intention prediction (No-LC, Left-LC, Right-LC). The method is evaluated on the highD and exiD datasets at 1s, 2s, and 3s prediction horizons, using location-based splits and imbalance-aware training procedures. The central claim is that the hybrid model consistently achieves higher macro-F1 than standalone LightGBM and Bi-LSTM baselines, especially on the ramp-rich exiD dataset.

Significance. If the reported results were reliable, the paper would make a modest empirical contribution: it would demonstrate that fusing learned temporal representations with structured physical features improves three-class lane-change intention prediction in both straight-highway and ramp scenarios, and it would add useful evidence on how class imbalance, history-window length, and prediction horizon interact. The paper also formulates a consistent labeling scheme for ramp areas and addresses imbalance with a multi-stage pipeline. However, the empirical basis is not currently established: the paper reports three mutually inconsistent sets of macro-F1 numbers, the history-window selection leaks test-set information, feature normalization appears to use dataset-wide statistics, and the Bi-LSTM training protocol is unspecified. These problems affect the central claim directly, so the significance of the contribution cannot be assessed as written.

major comments (5)
  1. [Abstract / §1 / Table 3] The central empirical claim is not supported by a single consistent result set. The arXiv-level abstract reports highD macro-F1 of 0.9514/0.9256/0.8872 and exiD macro-F1 of 0.9386/0.9070/0.8531; the full-text abstract and Section 1 contributions report highD 0.9562/0.9124/0.8345 and exiD 0.9247/0.8197/0.7605; Table 3 reports highD 0.9162/0.8847/0.8430 and exiD 0.7702/0.7233/0.6875. The spread is far beyond rounding (e.g., exiD T=1s differs by 0.168). Moreover, Table 3 does not support the claimed consistent advantage: on highD at T=1s the hybrid's Table 3 value (0.9162) is below both the stated LightGBM baseline (0.9359) and the Bi-LSTM baseline (0.9244) given in Section 1. The manuscript never reconciles these sets, so the reported evidence for the hybrid's superiority is internally contradictory.
  2. [§4.1] The history-window length W is selected using the held-out test set. The text states: "For each prediction horizon T, we evaluated every candidate window length W, recorded performance on both training and held-out data, and subsequently determined the window size that yielded the best generalization performance for the specified horizon." Because the held-out locations (highD 4–5, exiD 4–6) are used for model selection, the reported macro-F1 values are not unbiased estimates of generalization. The correct procedure would be nested cross-validation or fixing W a priori. This flaw affects all model comparisons and the "optimal performance" claims in Table 3.
  3. [§3.3, Eq. (14)] The normalization features z_i = (d_i - μ_i)/σ_i and s_i = d_i/μ_i use μ_i and σ_i derived from Figures 5 and 6, which are computed from dataset-wide distributions of lane-changing vehicles. If these statistics include test-location trajectories, then test-set information is used during feature construction, which would inflate the reported performance. The normalization statistics must be computed from the training split only; otherwise the experimental protocol is not leakage-free.
  4. [§2.3, §2.1, §3.3] The Bi-LSTM temporal encoder is a core component, but its training is unspecified: no loss function, no indication of whether the encoder is pretrained or trained end-to-end, no optimizer or hyperparameters, and no description of how the encoder labels are used. Equation (12) only defines pooling. The hybrid pipeline assumes that separately trained frozen embeddings remain useful when concatenated with handcrafted features for LightGBM, but no protocol is given. This makes the hybrid model irreproducible and the contribution of the temporal embeddings untestable.
  5. [§4.2, Table 3] No error bars, confidence intervals, or repeated-run statistics are reported. Macro-F1 differences between the hybrid and baselines are often small (e.g., highD T=2: 0.9124 vs 0.8916 Bi-LSTM in §1, or 0.8847 vs 0.8714 LightGBM in Table 3), so without variance estimates the claimed consistent advantage cannot be distinguished from noise. This is especially important given the test-set leakage in W selection.
minor comments (4)
  1. [§3.3, Eq. (14)] The placeholders for μ_i and σ_i in the text before Eq. (14) are missing, making the equation incomplete as printed.
  2. [Table 3] The label "Optimal performance of the Temporal Model" is ambiguous; the table should explicitly state that this refers to the proposed hybrid LightGBM+Bi-LSTM model, and it should include the baseline results in the same table for comparison.
  3. [§2.3, Eq. (12)] The pooling operation Pool(·) is left unspecified in the final implementation (mean, max, last, or attention). The choice should be reported because it affects the embedding and the results.
  4. [§3.3, Eq. (17)] There is a typo: "lane-cha0nge" should be "lane-change." There are also several missing spaces and incomplete equation placeholders throughout Section 3.3.

Circularity Check

2 steps flagged

Partial circularity: test-set-based window selection and label-derived 'physics-informed' features compromise the reported macro-F1 comparison.

specific steps
  1. fitted input called prediction [Section 4.1, Experimental setup (history-window selection)]
    "For each prediction horizon T, we evaluated every candidate window length W, recorded performance on both training and held-out data, and subsequently determined the window size that yielded the best generalization performance for the specified horizon."

    The held-out test split is used both to choose the history-window hyperparameter W and then to report the 'optimal' macro-F1 values in Table 3 and the Conclusion's claim of consistent superiority over baselines. Because the test data are the selection criterion, the reported held-out scores are an optimized quantity, not an independent prediction of generalization; the claimed advantage is therefore partly forced by fitting the evaluation set rather than by an unbiased comparison.

  2. self definitional [Section 3.3, Feature extraction, Equations (14)-(15) and Figures 5-6]
    "Figures 5 and 6 depict the average distance and distribution statistics between the lane-changing vehicle and adjacent vehicles ... For each distance d_i, we computed: z_i = (d_i − μ_i)/σ_i, s_i = d_i/μ_i ... A distance is considered safe if d_i > μ_i + 2σ_i ... Where μ_i and σ_i are the mean and standard deviation of Figure 5 and Figure 6, respectively."

    The normalization and safe-gap thresholds are defined using the spacing distributions of lane-changing vehicles, i.e., from the target behavior the model is supposed to predict. These dataset-level statistics enter the model's input features via Equations (14)-(15), with no stated restriction to the training split, so the 'physics-informed' predictor is partly constructed from the label distribution (including held-out locations) and is then said to predict lane-change intention from features that already encode lane-change statistics.

full rationale

The paper's central claim is an empirical model comparison, not a mathematical derivation, so most of it is not circular in the equation sense. However, two load-bearing steps exhibit partial circularity. First, Section 4.1 selects the history window W using the held-out test data and then reports performance on that same held-out data as the optimal generalization result; the claimed hybrid superiority is therefore partly an artifact of test-set selection. Second, the 'physics-informed' features in Equations (14)-(15) are normalized and thresholded using mean and standard deviation statistics computed from lane-changing vehicle spacing distributions (Figures 5-6) that are described as based on the full highD/exiD datasets; this encodes the target maneuver distribution into the input features, so the model's predictions are partially constructed from label-derived statistics. These issues make the reported macro-F1 values not fully independent predictions. Separately, the paper contains three mutually inconsistent sets of macro-F1 numbers (arXiv-level abstract vs. full-text abstract/contributions vs. Table 3), which is a serious correctness and reproducibility problem but not itself a circularity. The self-citations present are not load-bearing for the central derivation. A corrected protocol (nested cross-validation or fixed W, and training-only feature statistics) might still support the hybrid advantage, so the circularity is partial rather than total; thus score 6 rather than 8 or 10.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

All free parameters are fitted to the same datasets on which the model is evaluated, and one (W) is even selected on the held-out test set. The physics-informed features are handcrafted from domain knowledge rather than derived from first principles, so the contribution is an engineering recipe rather than a theory.

free parameters (5)
  • z-score normalization statistics \u03bc_i, \u03c3_i for gap distances = dataset-specific means/standard deviations (Figures 5-6)
    Equation (14) normalizes each neighbor distance by \u03bc_i and \u03c3_i taken from the dataset itself; if these statistics include test-location samples, features encode test-distribution information, and even if train-only, they are fitted to the data and not independently justified.
  • safe-gap threshold coefficient = 2 (multiples of \u03c3)
    Equation (15) defines safe gap as d_i > \u03bc_i + 2\u03c3_i; the 2\u03c3 cutoff is chosen by hand and determines the safe-gap indicator features.
  • inverse-frequency class weights = w_i = 1/f_c
    Equation (13) sets class weights from empirical class frequencies; these are fitted to the training distribution and directly affect the loss.
  • class-specific decision thresholds = optimized on validation set (values not reported)
    Section 3.2 states post-training threshold calibration on the validation set to maximize macro-F1; final threshold values are not reported, making the operating point hard to reproduce.
  • history window length W = selected per horizon on held-out data (exact values not reported)
    Section 4.1: W was chosen by evaluating on held-out test data, which is a test-set leakage in model selection and inflates reported performance.
axioms (4)
  • domain assumption The highD/exiD datasets provide accurate ground-truth lane changes via the stated labeling logic (0.2 m lateral crossing, 0.5 s monotonic drift, direction by laneId or lateral velocity)
    Section 2.1 defines start/end times and direction using hand-set thresholds; the entire label set rests on these thresholds being correct and consistent.
  • domain assumption Filtering out samples with multiple lane changes within the prediction horizon removes only ambiguous cases and does not bias the remaining evaluation
    Section 3.1 applies a strict maneuver-consistency rule that discards multi-maneuver trajectories, which changes the class distribution and may make the task easier than real-world multi-maneuver driving.
  • ad hoc to paper The Bi-LSTM encoder can be trained separately and its frozen embeddings remain useful when concatenated with handcrafted features for LightGBM
    The paper never specifies the Bi-LSTM training objective, loss, or whether it is trained end-to-end with the classifier; the fusion assumes the learned embedding transfers and adds value on top of the handcrafted features.
  • domain assumption Location-based splitting prevents trajectory overlap between train and test
    Section 4.1 splits by recording location, but the paper does not verify that the same physical vehicles do not reappear across locations or that the location split fully separates trajectories.

pith-pipeline@v1.3.0-alltime-deepseek · 12503 in / 11934 out tokens · 107162 ms · 2026-08-03T13:26:43.206291+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Evolutionary Physics-Informed Temporal Fusion for Lane-Change Intention Prediction." pith.science (2026). https://pith.science/paper/FKKOHUPA

@misc{pith2026251224075,
  author       = {Pith},
  title        = {Pith review of: Evolutionary Physics-Informed Temporal Fusion for Lane-Change Intention Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FKKOHUPA}},
  note         = {Machine review of arXiv:2512.24075}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Early lane-change intention prediction is essential for autonomous driving and ADAS, but it remains challenging because lane-changing behavior depends on evolving traffic risk, surrounding-vehicle interactions, and target-lane feasibility rather than only instantaneous vehicle states. This study proposes an evolutionary physics-informed temporal fusion framework for three-class lane-change intention prediction, including left lane change, right lane change, and no lane change. Instead of using static physics-informed variables alone, the proposed method derives temporal descriptors from conventional traffic signals, including risk evolution, gap persistence, counterfactual lane utility, interaction pressure gradient, maneuver feasibility, and intent consistency. These descriptors are fused with temporal embeddings learned from raw trajectory sequences through a sequence encoder, and the fused representation is used for final classification. Experiments are conducted on the highD and exiD datasets under 1\,s, 2\,s, and 3\,s prediction horizons. The proposed model achieves Macro F1-scores of 0.9514, 0.9256, and 0.8872 on highD, and 0.9386, 0.9070, and 0.8531 on exiD, respectively. The improvement is especially pronounced in exiD ramp-adjacent scenarios, indicating that temporal physical evolution is particularly useful in interaction-rich environments. These results demonstrate that combining evolutionary physics-informed descriptors with learned temporal representations provides a more dynamic and interpretable solution for early lane-change intention prediction.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 4 canonical work pages

  1. [1]

    Factors affecting lane change crashes,

    M. Shawky, "Factors affecting lane change crashes," IATSS Res. 44(2), 155 –161 (2020). https://doi.org/10.1016/j.iatssr.2019.12.002

    work page doi:10.1016/j.iatssr.2019.12.002 2020

  2. [2]

    Integrated driving behavior modeling,

    T. Toledo, H. N. Koutsopoulos, and M. Ben -Akiva, “Integrated driving behavior modeling,” Transportation Research Part C, 15(2), 96–112 (2007). https://doi.org/10.1016/j.trc.2007.02.002

    work page doi:10.1016/j.trc.2007.02.002 2007

  3. [3]

    General lane -changing model MOBIL for car -following models,

    A. Kesting, M. Treiber, and D. Helbing, “General lane -changing model MOBIL for car -following models,” Transportation Research Record, 1999(1), 86–94 (2007). https://doi.org/10.3141/1999-10

    doi:10.3141/1999-10 1999

  4. [4]

    Recent developments and research needs in modeling lane -changing behavior,

    Z. Zheng, “Recent developments and research needs in modeling lane -changing behavior,” Transportation Research Part B, 60, 16–32 (2014). https://doi.org/10.1016/j.trb.2013.11.009

    doi:10.1016/j.trb.2013.11.009 2014

  5. [5]

    Online inference of lane changing events for connected and automated vehicle applications with analytical logistic diffusion stochastic differential equation,

    H. Qi, C. Chen, X. Hu, and J. Zhang, “Online inference of lane changing events for connected and automated vehicle applications with analytical logistic diffusion stochastic differential equation,” Transportation Research Part C, 144, 103874 (2022). https://doi.org/10.1016/j.trc.2022.103874

    arXiv 2022

  6. [6]

    Experimental validation of lane-change intention prediction based on action recognition network,

    R. Izquierdo, I. Parra, D. Fernandez-Llorca, and M. A. Sotelo, “Experimental validation of lane-change intention prediction based on action recognition network,” IEEE ITSC (2019). https://doi.org/10.1109/ITSC.2019.8917331

    arXiv 2019

  7. [7]

    An improved learning-based LSTM approach for lane change intention prediction subject to imbalanced data,

    Q. Shi and H. Zhang, “An improved learning-based LSTM approach for lane change intention prediction subject to imbalanced data,” Transportation Research Part C , 133, 103414 (2021). https://doi.org/10.1016/j.trc.2021.103414

    arXiv 2021

  8. [8]

    MoCount: Motion-Based Repetitive Action Counting,

    R. Gu, S. Jia, Y. Ma, J. Zhong, J.-N. Hwang, and L. Li, “MoCount: Motion-Based Repetitive Action Counting,” in Proc. 33rd ACM Int. Conf. on Multimedia (ACM MM ’25), 2025, pp. 9026 –9034. https://doi.org/10.1145/3746027.3755857

    arXiv 2025

  9. [9]

    A comparison of headway and time to collision as safety indicators,

    K. Vogel, “A comparison of headway and time to collision as safety indicators,” Accident Analysis & Prevention, 35(3), 427–433 (2003). https://doi.org/10.1016/S0001-4575(02)00022-2

    work page doi:10.1016/s0001-4575(02)00022-2 2003

  10. [10]

    Vehicle -Level Safety Validation of AD/ADAS Systems via Extreme Value Analysis,

    P. Wu, S. Rahrovani, Z. Fei, D. Yang, S. Carlsson, and M. Törngren, “Vehicle -Level Safety Validation of AD/ADAS Systems via Extreme Value Analysis,” in Computer Safety, Reliability, and Security. SAFECOMP 2025 Workshops (LNCS 15955), 2025, pp. 437–452. https://doi.org/10.1007/978-3-032-02018-5_32

    work page doi:10.1007/978-3-032-02018-5_32 2025

  11. [11]

    Safety assured online guidance with airborne separation for urban air mobility operations in uncertain environments,

    P. Wu, X. Yang, P. Wei, and J. Chen, “Safety assured online guidance with airborne separation for urban air mobility operations in uncertain environments,” IEEE Trans. Intell. Transp. Syst., 23(10), 19413–19427 (2022). https://doi.org/10.1109/TITS.2022.3163657

    arXiv 2022

  12. [12]

    Cascading Credit Risk Assessment in Multiplex Supply Chain Networks,

    L. Chen, Y. Zou, P. Pan, and C. H. Chang, “Cascading Credit Risk Assessment in Multiplex Supply Chain Networks,” Authorea Preprints (Jan. 2026), doi: 10.22541/au.176858311.10362606/v1. https://doi.org/10.22541/au.176858311.10362606/v1

    arXiv 2026

  13. [13]

    Human Motion Instruction Tuning,

    L. Li, S. Jia, J. Wang, Z. Jiang, F. Zhou, J. Dai, T. Zhang, Z. Wu, and J.-N. Hwang, “Human Motion Instruction Tuning,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2025, pp. 17582–17591. https://doi.org/10.1109/CVPR52734.2025.01638

    arXiv 2025

  14. [14]

    ChatMotion: A Multimodal Multi -Agent for Human Motion Analysis,

    L. Li, S. Jia, J. Wang, Z. An, J. Li, J. -N. Hwang, and S. Belongie, “ChatMotion: A Multimodal Multi -Agent for Human Motion Analysis,” arXiv preprint arXiv:2502.18180 (2025). https://arxiv.org/abs/2502.18180

    Pith/arXiv arXiv 2025

  15. [15]

    Gradient boosting machines, a tutorial,

    A. Natekin and A. Knoll, “Gradient boosting machines, a tutorial,” Frontiers in Neurorobotics , 7, 21 (2013). https://doi.org/10.3389/fnbot.2013.00021

    arXiv 2013

  16. [16]

    Reinforcement Learning Approach for Highway Lane- Changing: PPO-Based Strategy Design,

    Z. Ma, Y. Luo, Z. Zhang, A. Sun, Y. Yang, and H. Liu, “Reinforcement Learning Approach for Highway Lane- Changing: PPO-Based Strategy Design,” in Proc. 10th Int. Conf. Electronic Technology and Information Science (ICETIS), 2025, pp. 298–301. https://doi.org/10.1109/ICETIS66286.2025.11144414

    arXiv 2025

  17. [17]

    Energy -Constrained Motion Planning and Scheduling for Autonomous Robots in Complex Environments,

    Z. Ma, A. Sun, Z. Zhang, Y. Yang, Z. Gao, and H. Liu, “Energy -Constrained Motion Planning and Scheduling for Autonomous Robots in Complex Environments,” in Proc. 5th Int. Conf. Advanced Algorithms and Neural Networks (AANN), 2025, pp. 591–594. https://doi.org/10.1109/AANN66429.2025.11257590

    arXiv 2025

  18. [18]

    Predicting asphalt pavement friction by using a texture-based image indicator,

    B. Lu, Z. Lu, Y. Qi, H. Guo, T. Sun, and Z. Zhao, “Predicting asphalt pavement friction by using a texture-based image indicator,” Lubricants, vol. 13, no. 8, p. 341, 2025

    2025

  19. [19]

    Depth -aware super -resolution via distance -adaptive variational formulation,

    T. Guo, B. Lu, F. Wang, and Z. Lu, “Depth -aware super -resolution via distance -adaptive variational formulation,” Journal of Electronic Imaging, vol. 34, no. 5, pp. 053018 –053018, 2025

    2025

  20. [20]

    ProSGNeRF: Progressive Dynamic Neural Scene Graph with Frequency Modulated Auto-Encoder in Urban Scenes,

    T. Deng, S. Liu, X. Wang, Y. Liu, D. Wang, and W. Chen, “ProSGNeRF: Progressive Dynamic Neural Scene Graph with Frequency Modulated Auto-Encoder in Urban Scenes,” arXiv preprint arXiv:2312.09076, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    GaussianDWM: 3D Gaussian Driving World Model for Unified Scene Understanding and Multi -Modal Generation,

    T. Deng, X. Chen, Y. Chen, Q. Chen, Y. Xu, L. Yang, L. Xu, Y. Zhang, B. Zhang, W. Huang, and H. Wang, “GaussianDWM: 3D Gaussian Driving World Model for Unified Scene Understanding and Multi -Modal Generation,” arXiv preprint arXiv:2512.23180, 2025

    Pith/arXiv arXiv 2025